Statistics › Section F · The normal distribution › Topic F.3

Choosing and checking a distribution

Every distribution in this course was built for a story. Real data do not announce their story - you read it from the context and the numbers, choose a model, then check that the model deserved your trust. This synthesis topic collects the whole family into one decision guide.

Which model?

Every distribution met so far was built for a story: counting successes, waiting for events, spreading chance over an interval. Real data do not announce their story - you read it from the context and the numbers, then choose. This synthesis topic collects the whole family into one decision guide, adds the checking habits that keep an honest modeller out of trouble, and shows what to do when no standard shape fits.

The story matcher

Match the story first, then the formulas follow. Counting successes in a fixed number of independent, same-\(p\) trials: binomial. Counting trials until the first success: geometric. Drawing without replacement from a small two-group population: hypergeometric. Counting events in a fixed interval at a known rate: Poisson. Waiting between those events: exponential. Values equally likely across a window: uniform. Measured quantities clustering symmetrically: normal. Two questions usually settle it: is the variable a count or a measurement, and is the number of trials fixed or open-ended?

Signature numbers

Each model leaves fingerprints in summary statistics. Binomial: mean \(np\), spread \(\sqrt{np(1-p)}\), values capped at \(n\). Poisson: mean and variance roughly equal. Exponential: mean and SD equal, with strong right skew. Uniform: mean at the window's midpoint, range spanning the window. Normal: symmetric, mean near median, nearly all values within 3 SDs. Compare a data set's mean, variance and shape against the candidate's signature - agreement is evidence, and a loud mismatch, like a variance far above the mean for counts, rules the model out.

Checking with pictures

Numbers check; pictures confirm. A histogram should show the shape the model claims: a bell, a flat rectangle, a peak near zero with a decaying tail. For a normal model specifically, the normal probability plot is the sharper tool: normal data fall close to a straight line, curvature reveals skew, and isolated points at the ends flag outliers. Plot before you compute - a few seconds of looking prevents pages of confidently wrong arithmetic.

Worked example: the cafe's arrivals

A cafe records customer counts per 10-minute period: mean 4.1, variance 4.3, values running from 0 to 11. 1. The story: customers arriving at random - events per interval. 2. The model: Poisson. 3. The signature: Poisson counts have variance near the mean, and 4.3 sits close to 4.1 - supported. 4. A bonus prediction: the gaps between arrivals should be exponential with mean \(\frac{10}{4.1} \approx 2.4\) minutes, a check the cafe could run with a stopwatch. 5. Check: the histogram should peak near 4 with a decaying right tail, the Poisson silhouette.

Worked example: four vignettes

Name each model. a) Free throws: 20 attempts, each scoring 70%, count the makes - binomial with \(n = 20\), \(p = 0.7\): fixed, independent, constant-\(p\) trials. b) A bus known to arrive within 15 minutes, equally likely anywhere in that window - uniform on \([0, 15]\). c) Scratches on a phone screen after a year - a count of small random events in a fixed interval - Poisson, with \(\lambda\) near the observed average. d) Adult heights - a measurement clustering symmetrically around a centre - normal. Check: each vignette answers the matcher's two questions - count or measurement, fixed or open-ended - before any formula is touched.

When the bell fails: skew and heavy tails

Right-skewed data - incomes, reaction times, page loads - fail every normality check, but all is not lost: taking logarithms often restores symmetry. Analysing \(\ln(\text{income})\) instead of income frequently makes normal-based methods appropriate, and results transform back for interpretation: small differences in natural logs read approximately as percentage changes. A different failure needs different care: quantities like stock returns and insurance claims carry heavier tails than the bell allows - extreme values occur far more often than the normal predicts - and forcing the normal model there underestimates the chance of disaster, a costly error in financial risk models. Diagnose by looking: if the biggest few observations tower past what 3 SDs allows, the bell is the wrong instrument. A transformation is itself a model choice, and it deserves the same shape checks as any model.

Mixtures

A histogram with two peaks often means two groups mixed into one data set - finishing times from a 5 km and a 10 km race pooled, response times of experts and novices combined. Two groups only show two peaks when their centres sit far apart compared with their spreads: pooled adult heights of men and women, a classic textbook example, usually show a single broad peak. The right move is usually to separate the groups and model each; the wrong move is to summarise the mixture with a single mean, a number that may describe nobody in either group. Bimodality is information: it says a lurking variable splits the data, and the split is the story worth telling.

Position without shape: percentiles

Percentiles describe where a value stands without assuming any distribution at all - which is why growth charts, standardised-test reports and medical reference ranges all use them. A child at the 80th percentile for height is taller than about 80% of comparable children, whatever the curve's shape. When no standard model fits, or when communicating to a non-technical audience, percentiles keep the description honest: shape-free, but concrete.

Models are approximations

George Box's famous dictum: all models are wrong, but some are useful. A distribution is a simplified description that earns its keep by making calculation and prediction possible - never by being exactly true. Judge a model by whether it is good enough for the decision at hand, check it against fresh data, and replace it without sentiment when the checks fail. The modeller's honesty lives in the checking habits of this topic, not in the elegance of the formulas.

Common mistakes to avoid

First, choosing a model by familiarity rather than by story. Second, assuming normality without a look at the shape. Third, ignoring heavy tails in risk calculations, which makes extreme losses look far rarer than they are. Fourth, summarising a bimodal mixture with one mean. Fifth, forgetting to transform back after a log analysis, leaving conclusions in log units. Sixth, treating a supported model as a proven one: agreement between mean and variance is evidence, not a certificate.

Key takeaways

Match the story to the model: binomial (fixed independent trials), geometric (trials to first success), hypergeometric (without-replacement draws from two groups), Poisson (events per interval), exponential (gaps between events), uniform (flat chance), normal (symmetric measurements). Check the signature numbers - Poisson's mean near variance, exponential's mean equal to SD, uniform's midpoint mean - and the pictures: histogram and probability plot. Skew can be tamed with logs; heavy tails and mixtures demand different tools. Percentiles describe position without shape. And Box's dictum governs it all: models are useful approximations, kept honest by checking.

Flashcards

Flashcard

Story matching

Read the scenario, then name the model

The modeller's first question.

Test your understanding

Question 1 of 5

The number of typos in a 50-page chapter is best modelled by:

Question 2 of 5

A player takes 20 free throws, each with a 70% chance of scoring. The number made is best modelled by:

Question 3 of 5

Count data have mean 4.1 and variance 4.3. This evidence most supports which model?

Question 4 of 5

Arrivals are Poisson with a mean of 4.1 per 10 minutes. What should the gaps between arrivals look like?

Question 5 of 5

Stock returns have heavier tails than a normal model allows. Fitting a normal curve to them tends to:

← Previous topicNext topic →